Shopify and Vercel are rebuilding Hydrogen as an open, runtime-agnostic foundation. Hydrogen now dep…
WebMCP vs. MCP for Commerce: Where the Agent Boundary Belongs
Two protocols now compete for the same job: letting an AI agent act on a storefront. WebMCP runs the actuation inside the browser, on top of whatever the user already has open. MCP runs it against a server-side endpoint the agent calls directly. Both solve "how does an agent take action," but they answer a different question first: where does that action happen, client or server? For teams building agent-ready storefronts in 2026, that placement decision is now an architecture choice, not an implementation detail.
Two Layers, One Job: Actuation
WebMCP exposes tools directly in the browser context. An agent operating inside a session, a browser extension, or an agentic browser can call functions the page itself registers, add to cart, apply a filter, start checkout, using the DOM and client-side state that already exist. We covered the actuation principle itself in how WebMCP lets agents act inside the browser; this piece looks at the placement decision, not the mechanism.
MCP (Model Context Protocol), by contrast, defines server-side endpoints an agent calls independent of any browser session: product search, cart mutation, order status, all exposed as callable tools on your backend or a middleware layer. We looked at what agents actually do once they reach those endpoints in MCP endpoints and storefront agent actions. Multiple vendors are now converging on server-side MCP exposure, a trend we tracked across the ecosystem in agent-readable storefronts across vendors.
The Boundary Decision, Concretely
- Where actuation runs. WebMCP (browser): Inside the loaded page, client-side. MCP (server): On a dedicated backend/middleware endpoint.
- What it needs. WebMCP (browser): A page that registers tools in the DOM. MCP (server): An API surface independent of any open page.
- Session context. WebMCP (browser): Uses whatever the shopper already has open (cart, login). MCP (server): Needs the agent to authenticate and hold its own session.
- Governance. WebMCP (browser): Bound to the frontend's own rendering, brand, and access rules. MCP (server): Bound to backend permissions, rate limits, and API contracts.
- Best fit. WebMCP (browser): Assistive actions inside a live shopping session. MCP (server): Autonomous or scheduled agents acting without a browser present.
Neither layer replaces the other. An agent inside a browser session benefits from WebMCP's direct DOM access; an agent operating without any open page, a backend integration, a scheduled reorder job, needs a server-side MCP endpoint that does not depend on a rendered page existing at all.
Why the Frontend Still Owns This Decision
Whichever layer an agent uses, it still reads the storefront's structured data, product attributes, and brand-consistent markup to decide what "add this to cart" even means. WebMCP tool registration lives in the frontend code. Server-side MCP endpoints still need frontend-maintained product and content data to answer correctly. That is exactly the operating question our Agentic Frontend Management Platform is built around: someone has to keep both actuation surfaces consistent, on-brand, and current, regardless of which protocol an individual agent picks.
What This Means for Merchants
- If your commerce team is building for in-session shopping assistants: start with WebMCP tool registration on your highest-intent pages.
- If you are exposing catalog, cart, or order data to autonomous agents without a browser present: build the MCP server endpoints first.
- Most storefronts will need both within 12 to 18 months. Plan the data layer once, expose it twice.
FAQ
Is WebMCP a replacement for MCP? No. WebMCP handles browser-side actuation for agents operating inside an open page or session. MCP handles server-side actuation for agents that call your backend directly, with or without a browser present.
Which one should merchants build first? It depends on where your agent traffic originates today. In-session shopping assistants favor WebMCP; backend integrations and autonomous agents favor MCP. Most teams end up needing both.
Does this change how storefronts should be architected? Yes, both actuation surfaces depend on the same underlying frontend data layer, structured product data, consistent markup, and brand governance, staying accurate. That is frontend management work, not a one-time protocol integration.
Next Steps
If you want a concrete assessment of where your storefront stands on browser versus server agent actuation, talk to us.
CTA: Check how agent-ready your storefront architecture is
More from the Laioutr Platform
About the author: Marcel Thiesies is CEO & Co-Founder of Laioutr. He writes about frontend architecture, agentic commerce, and building composable storefronts without replatforming risk.